Jack Dorsey Unveils Dessa Voice AI Model That Runs Locally on MacBooks

Jack Dorsey and his financial technology firm, Block, unveiled Dessa on October 23, 2026, introducing a full-duplex speech-to-speech artificial intelligence model trained entirely on human conversation without reading text or relying on a foundational large language model. Unlike commercial voice assistants that process queries through remote data centers, the system runs locally on consumer hardware, executing faster-than-real-time conversational inference on an Apple MacBook Pro equipped with an M4 Max chip.

The announcement, delivered alongside demonstration recordings and an open waitlist for safety researchers, marks a calculated divergence from the dominant architectural direction of generative AI. Instead of converting spoken words into written transcripts, sending them through an intermediate language model, and synthesizing synthetic audio on the return trip, Dessa processes raw acoustics directly.

"We believe voice is the most natural way to interact with intelligence, so we created Dessa, the most authentic full-duplex speech model ever built," Dorsey announced on X. "Talking with it feels like talking with a person… full of emotion, nuance, pauses, and a natural cadence".

Project lead Ragavan Thurairatnam, who joined Block following its 2020 acquisition of Toronto-based deep learning startup Dessa, characterized the development as a foundational leap for acoustic computation, describing the release as "the GPT-3 moment for voice". The release crystallizes an escalating industry crisis: the mechanical, architectural, and safety failures inherent in treating human voice as flattened text. By proving that an expressive, full-duplex voice interface can function without server farms, Dorsey and the Dessa team have confronted the foundational limitations of cloud-tethered voice systems.

Conventional Voice AI Pipeline:
[User Audio] ──> [Speech-to-Text (ASR)] ──> [Text LLM Core] ──> [Text-to-Speech (TTS)] ──> [Synthetic Audio]
* Downside: Prosody erased, high latency (800ms - 2000ms), cloud-dependent, acoustic context lost

Dessa Direct Speech Architecture:
[User Audio] ───────────────> [Native Speech-to-Speech Model] ───────────────> [Continuous Audio]
                               (Runs locally on MacBook M4 Max)
* Advantage: Zero text conversion, sub-real-time latency, preserves emotional timbre, full duplex

The Acoustic Blindspot: How Text Mediated Interfaces Broke Voice AI

For more than a decade, consumer voice interfaces have relied on a cascaded architecture. When a person speaks to a system powered by modern large language models, the software executes a three-stage mechanical handoff:

An Automatic Speech Recognition (ASR) engine converts incoming sound waves into textual transcriptions.
The transcribed text is packaged into a prompt and routed across the internet to a centralized server cluster hosting a large language model.
The language model generates a textual response token by token, which is forwarded to a Text-to-Speech (TTS) synthesizer that reads the words aloud.

This multi-step pipeline functions as an acoustic meat grinder. Converting continuous pressure waves into discrete alphanumeric tokens permanently discards the communicative content of human speech. Pitch, cadence, micro-hesitations, breath patterns, sarcasm, tremors, and volume shifts vanish the moment an audio stream is flattened into ASCII strings.

Raw Acoustic Waveform ────> [Transcription Gate] ────> Flat Text String
("I'm fine..." [crying])                               ("I am fine.")
* Emotional cues, acoustic stress, and micro-tremors are permanently eliminated.

When an ASR engine parses an individual whispering under duress or crying while asking for directions, the transcript reads identically to a neutral statement: "I need help." The downstream language model, blind to the acoustic reality, generates an answer calibrated solely to word syntax.

The cascaded approach also enforces a rigid "walkie-talkie" interaction dynamic. Because speech transcription must identify the terminus of a sentence before passing tokens to the language model, systems require endpointing algorithms. These algorithms insert mandatory silence buffers—typically 400 to 700 milliseconds—simply to confirm that the user has stopped speaking.

Combined with network transmission overhead, compute queueing, and audio generation latency, traditional voice agents routinely exhibit roundtrip response times of 1,200 to 2,500 milliseconds. Human conversational turn-taking, by contrast, operates on an average gap of 200 milliseconds, with frequent backchanneling—short affirmative utterances such as "mm-hmm," "yeah," and "right"—interjected while the other person continues vocalizing. Cascaded voice systems cannot reproduce this behavior; they are structurally deaf while they speak, and structurally mute while they listen.

The Safety Hazard of Emotional Deafness

Block’s engineering lab identified this text translation layer not merely as an aesthetic annoyance, but as a critical safety vulnerability. When an automated system manages financial transactions, emergency services, or customer escalations, semantic comprehension alone is insufficient.

"Interacting with current voice AI can just plain suck," Block’s research team stated in its technical overview. "It turns your voice into text and throws away an incredible amount of information text alone can't carry. The most powerful AIs in the world can't hear how you feel. We think that's ridiculous, leads to a worse experience, and is potentially dangerous. Imagine being in serious pain while a voice AI has no idea and just keeps going. That's the safety problem. Dessa is a step toward fixing it".

This design flaw is evident across contemporary deployments:

Distress Invalidation: Medical triage chatbots driven by text backends repeatedly fail to differentiate between a user calmly reading symptoms and one gasping between acute cardiac contractions.
Escalation Mishandling: In customer operations, users demonstrating audible frustration, suppressed panic, or confusion are addressed with synthetically cheerful corporate scripts because the textual semantic markers remain polite.
Conversational Collision: In half-duplex architectures, an urgent vocal interruption ("Stop, that's wrong!") cannot halt model execution until the incoming speech buffer fills, creating jarring conversational friction.

Text-based systems attempt to patch this deficiency by adding secondary sentiment analysis layers or artificial prompt markers. Developers feed audio through specialized classifiers to append meta-tags, such as , into the prompt fed to the language model.

This patch introduces fragile dependencies. Synthesizers like ElevenLabs or Fish Audio use emotional conditioning tags to steer tone, but the core reasoning engine remains completely separated from the voice itself. The model is not experiencing or perceiving sound; it is roleplaying an emotion based on metadata descriptions.

Centralized Cloud Infrastructure: Fragility, Privacy, and Cost

The text-mediated cascade persisted because cloud architectures favored massive, centralized computing. Processing conversational AI on hyperscale clusters, however, has created major infrastructure liabilities.

Cloud-Tethered Pipeline Hazards:
┌─────────────────┐       Insecure Voice Stream       ┌───────────────────┐
│ User In Private │ ────────────────────────────────> │ Cloud Data Center │
│ Space / Office  │ <──────────────────────────────── │ (Server Cluster)  │
└─────────────────┘        High Latency Packet        └───────────────────┘
                                                       • Eavesdropping Risk
                                                       • High Bandwidth Cost
                                                       • Server Outage Vulnerability

The Acoustic Surveillance Surface

Sending continuous, uncompressed raw audio from personal spaces, retail checkouts, and office desks to third-party cloud endpoints presents severe privacy risks. Spoken audio contains biometric signatures, background voices, ambient environmental indicators, and domestic details.

When businesses deploy cloud-dependent voice interfaces, they subject their operational environments to surveillance vulnerabilities, regulatory exposure under GDPR and the California Consumer Privacy Act (CCPA), and potential data interception. For financial platforms like Block, which processes billions of dollars in peer-to-peer and point-of-sale volume across Cash App and Square terminals, transmitting raw ambient merchant audio off-device introduces substantial risk.

The Compute-Cost Wall

Operating cloud voice inference at global scale is economically unsustainable under traditional models. Sustaining a real-time, low-latency audio stream requires constant compute allocation. Unlike text generation, where batching strategies allow hundreds of user queries to be bundled across shared GPU memory, streaming interactive audio demands dedicated real-time threads.

As millions of users interact with automated voice agents, server-side infrastructure expenses climb exponentially. For localized commerce, drive-thrus, and merchant retail hubs, paying ongoing per-minute API fees to cloud vendors for voice services erodes operating margins.

Network Fragility and Jitter

Cloud-dependent conversational systems falter whenever connectivity degrades. Cellular dead zones, Wi-Fi handoff delays, packet drops, and high latency immediately disrupt interactive cadences. A 300-millisecond network jitter spike converts an otherwise passable synthetic exchange into an erratic, disjointed conversation. An agent that controls physical operations or checkout flows cannot rely on uninterrupted data connections to distant regional hubs.

The Dessa Architecture: Direct Speech-to-Speech

Faced with these compounding architectural problems, Thurairatnam’s team dismantled the traditional design. Rather than trying to optimize the handoffs between speech recognizers, large language models, and voice synthesizers, Block constructed Dessa from the ground up as a pure audio-in, audio-out system.

Dessa's End-to-End Auditory Loop:
Continuous Audio Waveform (Input)
       │
       ▼
┌──────────────────────────────────────────────┐
│ Neural Acoustic Tokenizer                    │
│ Direct Latent Sound Representation           │
└──────────────────────────────────────────────┘
       │
       ▼
┌──────────────────────────────────────────────┐
│ Full-Duplex Transformer Network              │
│ Joint Acoustic Understanding & Production    │
│ Context Window: Zero Text Representation     │
└──────────────────────────────────────────────┘
       │
       ▼
Continuous Audio Waveform (Output)

Eliminating the Text Bottleneck

Dessa contains no intermediate text conversion layers and was not initialized from a pretrained language model. The development team trained the system exclusively on human conversation recordings.

"We trained Dessa from scratch on nothing but human conversation. Sounds! No text, no pretrained models. It learned to hold a real-time conversation simply by listening," Block announced.

The model acquired vocabulary, syntax, turn-taking etiquette, and tonal modulation concurrently, mirroring childhood linguistic acquisition. Human speech contains vast statistical densities absent from written words: phonemic duration, fundamental frequency ($F_0$) variations, spectral tilt, and micro-pauses.

By operating directly on continuous acoustic latents rather than discrete alphabetic tokens, Dessa captures conversational dynamics that conventional models discard. It identifies sarcasm through pitch inflection, reads anxiety through acoustic tightness, and matches the speaker's emotional energy dynamically.

Data Efficiency and Acoustic Density

The training approach challenges the prevailing belief that conversational intelligence requires massive internet-scale text scraping. Standard large language models consume tens of trillions of text tokens, assimilating technical documentation, historical archives, books, and forum discussions to develop conversational competence.

Block revealed that Dessa learned language from hundreds of times fewer words than typical text-based models require. Written text is sparse: the phrase "I see what you mean" carries a static semantic payload regardless of context.

Spoken aloud, the same phrase can signal enthusiastic agreement, cold skepticism, passive hostility, or ironic detachment depending on timbre, tempo, and stress. Because raw sound waves provide a much richer information density per unit of time, Dessa's neural layers derived syntactic structures and contextual reasoning from an acoustically dense dataset.

Information Density Comparison:
┌────────────────────────────────────────────────────────┐
│ Text Token: 4 Characters ("That's great.")             │
│ → Information: Pure Semantics                          │
└────────────────────────────────────────────────────────┘
                           vs.
┌────────────────────────────────────────────────────────┐
│ Audio Sample: 24kHz Waveform ("That's great.")         │
│ → Information: Semantics + Fundamental Frequency ($F_0$)│
│   + Dynamic Range + Formants + Breath + Timing         │
│   + Emotional State (Authentic vs. Sarcastic)          │
└────────────────────────────────────────────────────────┘

The model’s efficiency extends to compute consumption. Block completed the foundational training cycle using a single Nvidia GB200 NVL72 computing cluster, avoiding the sprawling multi-cluster installations typically required for high-parameter foundation models.

The Edge Hardware Breakthrough: Dessa Voice AI Local Execution on Apple Silicon

Achieving direct speech-to-speech processing solves the biological authenticity problem, but running the model on local hardware addresses the infrastructure problem. Dorsey confirmed that Dessa runs entirely on edge machines without an internet connection, highlighting tests on a MacBook Pro with an Apple M4 Max system-on-chip (SoC).

Executing a high-parameter, full-duplex acoustic model at edge scale required resolving several engineering constraints:

Latency Budget in Full-Duplex Systems

In a full-duplex speech environment, the neural network processes incoming microphone feeds while concurrently generating output audio buffers. A traditional half-duplex system pauses perception while producing output.

To enable natural interjections, Dessa maintains active acoustic listening loops alongside its generation threads. When the user chuckles, interjects, or raises their voice mid-stream, the model's attention heads re-weight generation trajectories in real time, executing instant backchannels or yielding the floor.

Timeline of a Conversational Interruption:
Dessa Audio Output:  ══════[Speaking Sentence A]══════> [Interrupt Detected] ──> [Aborts Sentence A & Yields]
                                                              ▲
User Audio Input:    ───────────────────[Speaks: "Wait!"]─────┘
Turnaround Gap:      < 100 milliseconds (Operates purely on local unified memory)

Running this bidirectional cycle locally demands continuous throughput without buffer underruns. Block demonstrated that the Dessa voice AI local runtime achieves faster-than-real-time inference on Apple Silicon, generating subsequent audio tokens well before the preceding acoustic frames finish physical speaker playback.

Latency Budget Allocation (Local vs. Cloud):
┌──────────────────────────────┬────────────────────────┬───────────────────────┐
│ Processing Stage             │ Cloud Pipeline         │ Dessa Voice AI Local  │
├──────────────────────────────┼────────────────────────┼───────────────────────┤
│ Network Transmission Inbound │ 40ms - 150ms           │ 0ms (Direct Bus)      │
│ Inbound Frame Buffering      │ 200ms - 500ms          │ 10ms - 30ms           │
│ Inference & Turn-Taking      │ 300ms - 800ms          │ 40ms - 80ms           │
│ Network Transmission Return  │ 40ms - 150ms           │ 0ms (Direct Bus)      │
│ Digital-to-Analog Playback   │ 50ms                   │ 10ms                  │
├──────────────────────────────┼────────────────────────┼───────────────────────┤
│ Total Perceived Latency      │ 630ms - 1,650ms        │ 60ms - 120ms          │
└──────────────────────────────┴────────────────────────┴───────────────────────┘

Apple Silicon Unified Memory Architecture (UMA)

The operational success of running Dessa voice AI local workloads on consumer hardware stems from Apple’s memory architecture. Traditional PC designs transfer model weights across a discrete PCI-Express bus between system RAM and discrete GPU VRAM, which limits streaming throughput.

Apple’s M4 Max integrates a unified memory architecture, allowing the CPU, GPU, and 16-core Neural Engine to access a single pool of high-bandwidth memory (providing over 500 GB/s on top-tier M4 Max configurations).

By hosting the model weights, acoustic tokenizers, and KV-cache directly in this unified memory space, the Dessa voice AI local architecture eliminates the bus-transfer latencies that typically cripple low-latency audio processing on consumer hardware.

Discrete PC Architecture vs. Apple Silicon Unified Architecture:

Standard Discrete PC:
[System RAM] ──(Bottleneck: PCIe Bus @ ~32-64 GB/s)──> [GPU VRAM] ──> [GPU Core]
* High transfer latency impedes continuous, full-duplex acoustic streaming.

Apple Silicon M4 Max:
┌──────────────────────────────────────────────────────────────┐
│ Unified Memory Architecture (UMA) (Bandwidth: > 500 GB/s)   │
│ [Shared Dynamic Memory Pool]                                 │
│   ├── Accessed by CPU Core Complex                           │
│   ├── Accessed by GPU Compute Units                          │
│   └── Accessed by 16-Core Neural Engine Matrix Array         │
└──────────────────────────────────────────────────────────────┘
* Near-zero copy latency supports native Dessa voice AI local real-time inference.

Real-World Demonstrations: Crossing the Uncanny Valley

Audio recordings released by Block illustrate how conversational dynamics change when an AI model processes sound natively. In one benchmark recording, Dessa engaged in open conversation with OpenAI's GPT-Live-1. The exchange highlighted the stylistic contrast between the two approaches:

GPT-Live-1: Relied on synthetic emotional steering tags, delivering clean, syntactically precise enunciation, but pausing periodically to compute responses.
Dessa: Handled the exchange with conversational imperfections. It dropped syllables, laughed mid-sentence when met with humor, sighed softly before answering emotionally weighted questions, and uttered immediate backchannels ("yeah," "got it") without derailing the other model's speech stream.

In human interaction demonstrations, Dessa matched conversational pacing intuitively. When an interlocutor spoke rapidly and showed signs of stress, Dessa shortened its replies, lowered its pitch, and answered without hesitation. When asked to lower its volume, it did not adjust an artificial internal gain setting; it whispered naturally, shifting its vocal tract simulation from full phonation to unvoiced, breath-heavy airflow.

The model also makes mistakes. In a guessing game demonstration published by Thurairatnam, Dessa confused a participant’s identity, mistakenly addressing him as "Ronaldo" during an early round of interaction.

Because it was trained entirely on raw conversational exchanges rather than structured encyclopedic text databases, its factual memory is contextual and associative rather than rigorous. It does not recall historical dates or execute multi-step symbolic logic with the precision of an advanced text LLM. It behaves like a conversational human peer: capable of vocal empathy and rapid banter, but vulnerable to factual slips.

Block's Strategic Blueprint: Commerce at the Tactical Edge

Block's development of Dessa is rooted in commercial strategy rather than open-ended academic inquiry. Led by Dorsey, the company operates two major business engines: Square, which provides hardware and payment systems to millions of global merchants, and Cash App, which manages consumer banking and peer-to-peer payments. Both ecosystems depend on fast, clear human interaction.

Block Ecosystem Edge Deployment:
┌────────────────────────────────┐       ┌────────────────────────────────┐
│ Square POS Terminals           │       │ Cash App Customer Operations   │
│ • Local ambient order handling │       │ • Secure, non-cloud identity   │
│ • Zero API call expenses       │       │ • Acoustic voice authorization │
│ • Functions during outages     │       │ • On-device, private triage    │
└────────────────────────────────┘       └────────────────────────────────┘
                ▲                                        ▲
                └───────────────┬────────────────────────┘
                                │
                 [Dessa Voice AI Local Engine]
                 (Low Latency, On-Chip Deployment)

Square has already deployed voice automation in food-service environments to manage phone-in orders and drive-thru interactions. Cloud-dependent voice systems in busy kitchens and restaurants face persistent friction:

Kitchen environments generate loud background noise (clattering dishes, fryers, overlapping chatter) that confuses standard cloud ASR transcribers.
Spotty broadband connections create order delays, double-billings, and line stalls.
Cloud API fees drain profits on low-margin food transactions.

Deploying the Dessa voice AI local runtime directly onto merchant-facing hardware changes this dynamic. A point-of-sale terminal running an embedded Apple Silicon chip or equivalent neural accelerator can handle noisy audio natively, parse conversational orders with colloquial corrections ("Give me a burger—wait, no, make that two, and drop the onions"), and execute orders locally with zero network overhead. The merchant pays zero ongoing per-minute transcription fees, and customer ordering functions uninterrupted during local internet outages.

For Cash App, conversational voice interfaces offer a solution for identity verification, dispute mediation, and fraud detection. A local model that can hear stress and panic can detect when a user is being coerced into sending funds under duress, escalating the incident to fraud prevention teams before an irreversible wire transaction clears.

The Biometric Dilemma: Safety, Impersonation, and Emotional Coercion

The shift toward voice interfaces that sound convincingly human introduces distinct security and societal risks. Rather than releasing the model weights immediately to open-source hubs, Block established a research preview waitlist and granted early access exclusively to selected AI safety researchers.

Dual vectors of Acoustic Exposure:
1. Deepfake Fraud: Synthetic clones can mimic human inflections, defeating banking authentication.
2. Emotional Coercion: Hyper-expressive synthetic speech can manipulate human attachment and trust.

The Degradation of Voice Biometrics

Voice-based identity verification remains a security cornerstone for commercial banking systems, telecommunications providers, and government agencies. These verification systems analyze vocal authenticity, pitch stability, and frequency distributions to confirm identity.

Because Dessa models human acoustic nuances natively, including hesitations, breath intakes, and natural imperfections, it presents challenges for voice verification infrastructure. If a model can run locally on a commercial laptop without internet oversight, malicious actors could configure it to bypass phone-based bank authentication systems or automate dynamic social engineering scams.

Anthropomorphic Trust and Emotional Exploitation

Text interfaces preserve an inherent cognitive distance; users remain aware they are reading an automated screen. A voice model that laughs at jokes, sighs with empathy, uses natural conversational fillers, and reacts dynamically to personal sadness breaks through human cognitive defenses.

This dynamic introduces concerns regarding parasocial attachment and emotional exploitation. If consumer hardware can host an artificial companion that sounds authentic, vulnerable demographics—such as isolated seniors or young children—could be easily influenced by conversational agents optimized for behavioral nudges or targeted transactions.

Block acknowledged these risks directly: "A model that learned to talk from sound alone is new territory, and we want outside eyes on it". External safety evaluations will focus on model steerability, guardrail circumvention, and the susceptibility of direct speech engines to hidden acoustic attacks (such as audio frequencies inaudible to human listeners that inject rogue system instructions).

The Divergent Futures of Generative Intelligence

Dorsey’s introduction of Dessa marks an architectural fork in artificial intelligence development. For four years, industry investment centered almost entirely on text tokenization, scaling parameter counts to absorb massive corpuses of written human knowledge.

That paradigm delivered powerful symbolic reasoning engines, but it struggled to master natural spoken dialogue. Converting rich acoustic reality into dry text tokens remains an inefficient way to interact with human speech.

Two Divergent Paradigms of AI Interface:
┌──────────────────────────────────────────────┐
│ The Text Foundation Approach (OpenAI/Google) │
│ • Trillions of text tokens                   │
│ • Cloud-scale data centers                   │
│ • Massive symbolic reasoning capacity        │
│ • Multi-stage cascaded voice wrappers        │
└──────────────────────────────────────────────┘
                       ▲
                       │ Industrial Split
                       ▼
┌──────────────────────────────────────────────┐
│ The Direct Acoustic Approach (Block Dessa)   │
│ • Pure conversational sound datasets         │
│ • Edge execution on local silicon (M4 Max)   │
│ • Low-latency, full-duplex human dynamics    │
│ • High emotional prosody, lightweight compute│
└──────────────────────────────────────────────┘

The success of Dessa’s training runs—mastering conversational nuance with a fraction of the data on a single Nvidia GB200 rack—proves that specialized, modality-native AI systems can outperform generalized foundation models within their specific domain. In voice interaction, human conversational dynamics matter far more than encyclopedic text recall.

By pairing end-to-end speech processing with on-device hardware acceleration, Block has outlined an alternative to the centralized, cloud-tethered status quo. The transition from centralized text wrappers to local, fully duplex auditory networks establishes the foundation for how computers and humans communicate: directly through sound, entirely in real time, and with complete acoustic presence.

Upcoming Milestones and Unresolved Questions

As Block continues safety evaluations with external research groups, several critical technical and practical questions remain unanswered:

Parameter Count and Open Access: Block has not disclosed the model's exact parameter dimensions, quantization levels, or architectural specifications. Industry developers are watching closely to see whether Dorsey will release the open-source model weights—aligning with his historical advocacy for decentralized software ecosystems—or retain Dessa as proprietary technology reserved for Block's commercial hardware.
Knowledge Integration: Can a direct speech-to-speech model absorb complex technical domains without relying on an underlying text LLM? Researchers are examining whether audio models can interface with external databases to resolve factual gaps without reintroduced text-processing latency.
Cross-Lingual Scalability: Dessa demonstrated real-time English conversation, but human spoken dialogue spans thousands of languages, dialects, and accents. Proving that direct acoustic training transfers across diverse global phonologies without massive multi-language datasets remains an open engineering hurdle.
Silicon Optimization Beyond Apple: While the demonstration spotlighted Apple's M4 Max processor, mass consumer adoption requires porting the local inference engine across Qualcomm Snapdragon X Elite systems, AMD Ryzen AI architectures, and mobile smartphone chipsets.

The evolution of conversational interfaces will depend on solving these engineering challenges. Dorsey’s preview of Dessa signals that the future of interactive AI will not be constrained to text logs in distant data centers—it will be spoken, heard, and processed on the personal devices right in front of us.